GPT-5.6 Benchmark Review: Sol, Terra, and Luna Across 12 Benchmarks

OpenAI shipped GPT-5.6 on July 9, 2026: three tiers (Sol, Terra, Luna), a 5x price spread from $1 to $5 per million input tokens. Sol leads Agents' Last Exam at 53.6%. Terra scores within 3 points on most benchmarks at half the cost. Luna matches last generation's flagship for one-fifth the price.

We ran all three through Stratix. Here is where each tier holds up and where it breaks down.

TL;DR

• GPT-5.6 Sol scores 53.6% on Agents' Last Exam, beating Claude Fable 5 by 13.1 points on long-horizon professional workflows.

• Terra scores within 3 points of Sol on most benchmarks at half the token cost ($2.50/$15 per 1M vs. $5/$30).

• Luna ($1/$6 per 1M tokens) matches GPT-5.5 on Terminal-Bench at 84.7% while costing roughly one-fifth as much.

• Sol's 1.05M-token context window holds: 91.5% on MRCR 8-needle recall at 256K-512K. Luna drops to 41.3% on the same test.

• "Ultra" multi-agent mode (Sol only) pushes BrowseComp to 92.2%. Token cost scales 4x.

• Terra handles development loops. Sol handles production gates. Luna handles volume.

Three Tiers, One Architecture

GPT-5.6 ships as one architecture distilled into three capability tiers, each with the same API surface and reasoning modes (none, low, medium, high, xhigh, max). Sol adds "ultra," which coordinates four parallel sub-agents.

The pricing structure rewards volume. Cached input tokens drop to $0.50/1M for Sol, and requests above 272K tokens jump to $10/$45. For teams running continuous evaluation loops, the cache discount changes the math significantly.

Coding: 1.2 Points Separate Sol and Terra on SWE-Bench Pro

On the Artificial Analysis Coding Agent Index, Sol with max reasoning scores 80, 2.8 points above Fable 5. Terra scores 77.4. Luna scores 74.6, above Opus 4.8's 72.5.

On SWE-Bench Pro, Sol completes tasks at 64.6% accuracy. Terra hits 63.4% with fewer tokens per completion. That 1.2-point gap costs 2x more per million tokens.

Terminal-Bench 2.1: Sol at 88.8%, Terra at 87.4%, Luna at 84.7%. Luna at $1 input lands within a point of GPT-5.5's 85.6% at one-fifth the price. For coding agents in CI/CD pipelines where cost-per-run compounds, Luna handles the volume while Sol handles the edge cases.

Knowledge Work and Computer Use

OSWorld 2.0 is where Sol pulls away: 62.6% vs. Terra's 50.2% and Luna's 45.6%. This is a computer-use benchmark, and Sol's ability to inspect rendered output (not just generate code) gives it a structural advantage. Opus 4.8 scores 54.8% here.

BrowseComp: Sol at 90.4%, Terra at 87.5%, Luna at 83.3%. The "ultra" mode pushes Sol to 92.2% by running four agents in parallel. Token cost scales linearly with agent count.

GDPval-AA v2 (document and presentation quality) is where Fable 5 still leads: 1,759.6 Elo vs. Sol's 1,747.8. Terra drops to 1,593 Elo. For document generation workflows, Sol competes with Fable 5. Terra does not.

Long-Context: Luna's Ceiling

All three tiers accept 1.05M tokens. Effective recall at that length is a different question.

MRCR v2 8-needle at 256K-512K: Sol 91.5%, Terra 89.6%, Luna 41.3%. At 512K-1M: Sol 73.8%, Terra 72.5%, Luna 41.3%.

Luna accepts long inputs. It does not use them. If your pipeline processes large documents or codebases, Terra is the minimum viable tier.

Cybersecurity: Sol's Clearest Win

This is where the tiering gap is widest. On ExploitBench, Sol scores 73.5% vs. Terra's 52.9% and Luna's 33.2%. On SEC-Bench Pro, Sol hits 71.2% vs. Terra's 57.7%. On Capture-the-Flag challenges, Sol scores 96.7%.

OpenAI is positioning GPT-5.6 as a defensive security tool, with its Trusted Access for Cyber program gating the most capable security features behind identity verification and hardware-backed passkeys. For security teams, Sol is the only tier worth evaluating.

Benchmarks Set the Shortlist. Your Production Data Sets the Decision.

Every score above was measured in controlled conditions. Your agent runs unsupervised, processes customer data, and makes decisions your compliance team has never reviewed.

A model scoring 80 on a coding index may score 45 on your specific deployment pipeline. Benchmarks narrow the field. Continuous evaluation on your tasks, with your data, against your quality bar, determines which tier you actually ship.

Which Tier Should You Use?

Luna ($1/$6): Default for high-volume, cost-sensitive workloads with short contexts. Classification, routing, simple summarization. Replaces GPT-5.5 at one-fifth the cost for most tasks.

Terra ($2.50/$15): Development loops, coding agents in CI/CD, document processing up to 512K tokens. Matches or beats GPT-5.5 on nearly every benchmark at half the price.

Sol ($5/$30): Production gates, security workflows, long-context analysis, computer use. The only tier where "ultra" is available. Reserve for tasks where the cost of a wrong answer exceeds the cost of the tokens.

Key Takeaways

• Run your own evaluations before committing to a tier. Benchmark scores compress real-world performance differences. A model scoring 3 points higher on SWE-Bench Pro may score 15 points higher on your specific codebase.

• Start with Terra, escalate to Sol only where accuracy gaps justify the 2x cost increase. Most teams over-buy on model capability.

• Luna's long-context collapse at 41.3% recall is a hard constraint, not a soft trade-off. Do not route long-document tasks to Luna regardless of cost pressure.

• Test the "ultra" multi-agent mode on your hardest tasks before dismissing it. The 4x token cost is steep, but for tasks with high error cost (security audits, financial analysis), the accuracy gain may justify it.

• Set up continuous evaluation comparing tiers on your production traffic. Model selection requires ongoing optimization as your production workloads evolve.

FAQ

How does GPT-5.6 Sol compare to Claude Fable 5 overall?

Sol leads on agentic workflows (Agents' Last Exam: 53.6% vs. 40.5%) and coding efficiency (fewer tokens, lower cost for comparable scores). Fable 5 leads on the Artificial Analysis Intelligence Index (59.9 vs. 58.9) and GDPval (1,759.6 vs. 1,747.8 Elo). Pick based on whether your workload is heavier on agent execution or document generation.

Is GPT-5.6 Terra a viable replacement for GPT-5.5?

Yes, for most workloads. Terra matches or exceeds GPT-5.5 on every major benchmark we tested while costing roughly half as much per token. The one exception is long-context recall above 512K tokens, where performance is comparable but not clearly better.

What is "ultra" mode and when should I use it?

Ultra coordinates four parallel sub-agents (configurable up to 16) to tackle complex tasks simultaneously. It is available only on Sol. Use it for tasks with high error cost and sufficient complexity to benefit from parallel exploration: security audits, multi-document analysis, and complex debugging sessions.

Can I use Luna for agent evaluation?

Luna works for evaluating simple, single-turn agent responses. For multi-turn agent trajectories or tasks requiring long-context understanding, Terra is the minimum viable tier due to Luna's context recall limitations.

What benchmarks did you use for this evaluation?

We evaluated across Agents' Last Exam, Artificial Analysis Coding Agent Index, Terminal-Bench 2.1, SWE-Bench Pro, DeepSWE, BrowseComp, OSWorld 2.0, GDPval-AA v2, MRCR v2, ExploitBench, SEC-Bench Pro, and GPQA Diamond. Full methodology is described below.

Methodology

Benchmark scores referenced in this article combine OpenAI's published results with independent evaluations from Artificial Analysis and our own Stratix evaluation runs. Where scores differ between sources, we note the source. All pricing is based on OpenAI's published API rates as of July 2026.

Evaluate GPT-5.6 on your own tasks. Stratix lets you compare Sol, Terra, and Luna against 200+ models on 52+ benchmarks, or bring your own evaluation datasets. Try Stratix free →